Upload
facultad-de-informatica-ucm
View
22
Download
2
Embed Size (px)
Citation preview
Universitat Politècnica de CatalunyaBARCELONA TECH
Madrid. March 30, 2017
Eduard Ayguadé
Dep. Arquitectura de Computadors, U. Politècnica de CatalunyaComputer Sciences Dep., Barcelona Supercomputing Center
OpenMP tasking model: from the standard to the classroom
ICT386 multiprocessor architecture, late 80s
1
4 x CPU modulei386/i387no private cache memory
4 x Memory module4 MB DRAM (100 ns)64 KB shared cache25 ns, direct-mapped
ICT386 multiprocessor architecture, late 80s
2
Crossbar 5x5Gate array design10 ns transfer
Intellligent memoryMemory mapped
1 MB DRAMFunctions: management,
synchronization andscheduling
Prog
ram
mab
leN
IC
And the wall has become higher and higher …
3
big.LITTLE Integrated GPU
Sock
et
S1 S2
GPUGDDR
GPUGDDR
DDR
DDR
FPGANVMHCA
CAPIPCIe
PCIePCIe
QPI
Node
Rank Name Site Computer TotalCores
Accelerator/Co-Proc.
CoresRmax (Gflops) Rpeak
(Gflops) Power (KW) Mflops/Watt
1Sunway TaihuLight
National Supercomputing Center in Wuxi
Sunway MPP, Sunway SW26010 260C 1.45GHz, Sunway 10649600 0 93014593,88 125435904 15371 6051,30
2Tianhe-2 (MilkyWay-2)
National Super Computer Center in Guangzhou
TH-IVB-FEP Cluster, Intel Xeon E5-2692 12C 2.200GHz, TH Express-2, Intel Xeon Phi 31S1P 3120000 2736000 33862700 54902400 17808 1901,54
3 TitanDOE/SC/Oak Ridge National Laboratory
Cray XK7, Opteron 6274 16C 2.200GHz, Cray Gemini interconnect, NVIDIA K20x 560640 261632 17590000 27112550 8209 2142,77
4 Sequoia DOE/NNSA/LLNLBlueGene/Q, Power BQC 16C 1.60 GHz, Custom 1572864 0 17173224 20132659,2 7890 2176,58
5 Cori DOE/SC/LBNL/NERSC
Cray XC40, Intel Xeon Phi 7250 68C 1.4GHz, Aries interconnect 622336 0 14014700 27880653 3939 3557,93
6Oakforest-PACS
Joint Center for Advanced High Performance Computing
PRIMERGY CX1640 M1, Intel Xeon Phi 7250 68C 1.4GHz, Intel Omni-Path 556104 0 13554600 24913459 2718 4985,69
7 K computer
RIKEN Advanced Institute for Computational Science (AICS)
K computer, SPARC64 VIIIfx 2.0GHz, Tofu interconnect 705024 0 10510000 11280384 12660 830,18
8 Piz Daint
Swiss National Supercomputing Centre (CSCS)
Cray XC50, Xeon E5-2690v3 12C 2.6GHz, Aries interconnect , NVIDIA Tesla P100 206720 170240 9779000 15987968 1312 7453,51
9 MiraDOE/SC/Argonne National Laboratory
BlueGene/Q, Power BQC 16C 1.60GHz, Custom 786432 0 8586612 10066330 3945 2176,58
10 Trinity DOE/NNSA/LANL/SNL
Cray XC40, Xeon E5-2698v3 16C 2.3GHz, Aries interconnect 301056 0 8100900 11078861 4232 1913,92
Top5 November 2016 list
1979 Pink Floyd’s The Wall album,Cray-1 delivering 80 Mflops, single processor
OpenMP evolution and OmpSs influenceRuntime opportunitiesArchitecture opportunitiesTeaching opportunities
Programmability
The vision
Plethora of architectures• Heterogeneity• Hierarchies
ISA Low-level interfaces
ISA/PM leaks and variability/faults/errors everywhere• Programmer exposed• Divergence between
mental models and actual behavior
New application domains
Programming models
Applications
Programmability andporting/inter-operability• Legacy codes• ”Legacy” mental models
and habits
APIs
New usage practices• Analytics and visualization• New data models• Interactive supercomputing, response
time
• Virtualized data centers• Cloud
Programming interfaces
Ana…
…lytics
9
The vision
10
Intelligent runtime, parallelization, distribution, interoperability
General purposeTask basedSingle address space
Program logicindependent of computing platform
Performance-and power-drivenoptimization/tuning
Resource management
ISA Low-level interfaces APIs
Runtime system Resource managementPerformance, power, energy… monitoring/insight
Applications
Programming model: high-level, clean, abstract interface
7
The OmpSs forerunner for OpenMP
Proposing ideas …– IWOMP workshop is the ideal scenario– Participation in the OpenMP language committee and sub-committees
Demonstrated with …– Prototype implementation: Mercurium compiler and Nanos runtime– Benchmark/application suite
Lots of discussions and final polishing
3.0 2008
+ Tasks + Taskdependences
+ Taskpriorities
+ Taskloopprototyping
+ Task reductions+ Dependences
on taskwait
+ Multidependences+ Commutative
+ Dependenceson taskloop
Today
3.1 2011
4.0 2013
4.5 2015
5.0 ?OmpSs
8
OmpSs influencing tasking model evolution (1)
The big change …
From loop-based model to task-based model– Dynamically allocated data structures
(lists, trees, …)– Recursion
task generation waits for child tasks to finish (3.0)
task-based model, nesting of tasks (3.0)
#pragma omp task firstprivate(list){ code block }
#pragma omp taskwait
task priorities (4.5)
priority(value)
task generation cut-off (3.1)
final(expr)
root
9
OmpSs influencing tasking model evolution (2)
Directionality of task arguments
Dependences between tasks computed at runtimeTasks executed as soon as all dependencesare satisfied
directionality for task arguments (4.0)
#pragma omp task depend(in: list, out: list, inout: list)
{ code block }
10
OmpSs influencing tasking model evolution (3)
Tasks become chunks of iterations of one or more associated loops
Dependences between tasks generated in the same or different taskloop construct under discussion
tasks out of loops and granularity control (4.5)
#pragma omp taskloop [num_tasks(value) | grainsize(value)]for-loops
#pragma omp taskloop block(2) grainsize(B, B)depend(in : block[i-1][j], block[i][j-1])depend(in : block[i+1][j], block[i][j+1])depend(out: block[i][j])
for (i=1; i<n i++)for (j=1; j<n;j++)
foo(i, j);
T T T T
T T T T
T T T T
T T T T
11
OmpSs influencing tasking model evolution (4)
Reduction operations on task and taskloop not yet supported, currently in discussion#pragma omp taskgroup reduction(+: res){while (node) {#pragma omp task in_reduction(+: res)res += node->value;
node = node->next;}
}
#pragma omp taskloop grainsize(BS) reduction(+: res)for (i = 0; i < N; ++i) {res += foo(v[i]);
}
12
OmpSs influencing tasking model evolution (5)
The number of task dependences have to be constant at compile time à multidependences proposal under discussion
for (i = 0; i < l.size(); ++i) {#pragma omp task depend(inout: l[i]) // T1{ ... }
}
#pragma omp task depend(in: {l[j], j=0:l.size()}) // T2for (i = 0; i < l.size(); ++i) {...
}
Defines an iterator with a range of values that only exists in the context of the directive. Equivalent to:depend(in: l[0], l[1], ..., l[l.size()-1])
T1 T1 T1 T1
T2
…
13
OmpSs influencing tasking model evolution (6)
New dependence type that relaxes the execution order of commutative tasks but keeping mutual exclusion between them
T1
T3
T2
T2
for (i = 0; i < l.size(); ++i) {#pragma omp task depend(out: l[i])// T1
}
for (i = 0; i < l.size(); ++i) {#pragma omp task shared(x) depend(in: l[i])
depend(inout: x)// T2
}
#pragma omp task shared(x) depend(in: x) // T3for (i = 0; i < l.size(); ++i) {...
}
T1T1 T1
T2
T2
14
OmpSs influencing tasking model evolution (6)
New dependence type that relaxes the execution order of commutative tasks but keeping mutual exclusion between them
T1
T3
T2
T2
for (i = 0; i < l.size(); ++i) {#pragma omp task depend(out: l[i])// T1
}
for (i = 0; i < l.size(); ++i) {#pragma omp task shared(x) depend(in: l[i])
depend(commutative: x)// T2
}
#pragma omp task shared(x) depend(in: x) // T3for (i = 0; i < l.size(); ++i) {...
}
T1T1 T1
T2
T2
Scheduling
New tasks
Dynamic task graph
Worker threads
Execute task
Nan
os++
runt
ime
…newTask(call0, data0);newTask(call1, data1);…waitTasks()…
Application binary Readytask queue
End task: dependence update
15
OmpSs: runtime opportunities (1)
Runtime parallelization– Compute data dependences among tasks à dynamic task graph– Dataflow execution of tasks à ready task queue– Look-ahead: task instantiation can proceed past non-ready tasks (with
opportunities to find distant parallelism)
16
OmpSs: runtime opportunities (2)
Criticallity-aware scheduler: managing heterogeneous SoC– Tasks in the critical path detected at runtime à execute in faster cores– Non-critical tasks executed in slower, more power-efficient cores– Work-stealing for load balancing
“Criticality-Aware Dynamic Task Scheduling on Heterogeneous Architectures”. K. Chronaki et al. ICS-2015
0,70,80,9
11,11,21,3
Cholesky Int. Hist QR Heat AVG
Spee
dup Original CATS
Scheduling
New tasks
Dynamic task graph
Worker threads
Execute task
Nan
os++
runt
ime
…newTask(call0, data0);newTask(call1, data1);…waitTasks()…
Application binary Readytask queue
End task: dependence update
17
OmpSs: runtime opportunities (3)
Runtime parallelizationDirectory with per-core arguments utilization– Locality-aware task scheduling strategies– Driving argument prefetching
Directory
18
OmpSs: runtime opportunities (4)
Management of address spaces and coherence– Directory@master: identifies for every region reference where data is– Per-device software cache: translate host to device address spaces– Use of device specific mechanisms to move data, if needed– Locality-aware scheduling when multiple devices exist
Scheduling
Dynamic task graph
Nan
os++
runt
ime
Readytask queue
End task: dependence update
Worker threads
Helper threads
Execute task
MIC threads
CUDA threads
Datarequests
Directory / Caches
“Self-Adaptive OmpSs Tasks in Heterogeneous Environments”. J. Planas et al., IPDPS 2013.
Example: nbody
#pragma omp target device(smp) copy_deps#pragma omp task depend(out: [size]out), in: [npart]part)void comp_force (int size, float time, int npart, Part* part, Particle *out, int gid);
#pragma omp target device(fpga) ndrange(1,size,128) copy_deps implements (comp_force)#pragma omp task depend(out: [size]out), in: [npart]part)__kernel void comp_force_opencl (int size, float time, int npart, __global Part* part,
__global Part* out, int gid);
#pragma omp target device(cuda) ndrange(1,size,128) copy_deps implements (comp_force)#pragma omp task depend(out: [size]out), in: [npart]part)__global__ void comp_force_cuda (int size, float time, int npar, Part* part, Particle *out, int gid);
void Particle_array_calculate_forces(Particle* input, Particle *output, int npart, float time) {for (int i = 0; i < npart; i += BS )
comp_force (BS, time, npart, input, &output[i], i);}
20
OmpSs: runtime opportunities (5)
Support for multiple implementations for same task– Versioning scheduler:
• Monitoring (learning) phase and scheduling phase• Most suitable implementation is selected each time depending on
monitoring, simple model and OmpSs worker’s work load
SMP W1
SMP W2
GPU W3Faster but
busy
Slower but early executor
Run task!
“Self-Adaptive OmpSs Tasks in Heterogeneous Environments”. J. Planas et al., IPDPS 2013.
21
OmpSs: runtime opportunities (6)
At cluster …– Task creation starts at master node, outermost tasks may be executed
on remote nodes and spawn inner tasks– Locality-aware scheduler, overlap transfers/computation, work stealing
Scheduling
Directory/Cache
Nano
s++
runt
ime
mas
ter n
ode
Worker threads
Helper threads
Execute task
Datarequests
Worker threads
Helper threads
Nano
s++
runt
ime
rem
ote
node
data in host, devices attached to node, and remote nodes
data in host, devices attached to node,
NIC NIC
“Productive Cluster Programming with OmpSs”. J. Bueno et al. EuroPar 2011.“Productive Programming of GPU Clusters with OmpSs”. J. Bueno et al., IPDPS 2012.
22
OmpSs: architecture opportunities (1)
Novel hybrid cache/local memory– Cache: irregular accesses– Local memory (energy efficient, less
coherence traffic): strided accesses
Runtime-assisted prefetching• Task inputs and outputs mapped to the LMs• Overlap with runtime• Double buffering between tasks
8.7% speedup, 14% reduction in power,20% reduction in network-on-chip traffic
C CL1 Cl
usterInterconn
ect
LML1LM
C CL1LM
L1LM
C CL1LM
L1LM
C CL1LM
L1LM
DRAM DRAM
L2
L3
“Runtime-Guided Management of Scratchpad Memories in Multicores.” Ll. Alvarez et al. ÇPACT-2015
OmpSs: architecture opportunities (3)
Task dependence manager to speedup task management and dependence tracking
Hardware support for reductions– Privatization of variables to speedup reduction à Significant overhead
when performing reductions on a vector or matrix block.– HW support to perform simple reduction operations (+, -, *, max, min)
in the last-level cache or in memory à In-memory computation
X. Tan et al. Performance Analysis of a Hardware Accelerator of Dependency Management for Task-based Dataflow Programming models. ISPASS 2016.X. Tan et al. General Purpose Task-Dependence Management Hardware for Task-based Dataflow Programming Models. IPDPS 2017.
J. Ciesko et al. Supporting Adaptive Privatization Techniques for Irregular Array Reductions in Task-parallel Programming Models. IWOMP 2016.
24
OmpSs: architecture opportunities (2)
CATA: criticallity-aware task acceleration– Runtime drives per-core DVFS reconfigurations meeting a global
power budget– Hardware Runtime Support Unit (RSU) to minimize overheads that
grow with the number of cores
E. Castillo et al., CATA: Criticality Aware Task Acceleration for Multicore Processors. IPDPS 2016.
80%90%
100%110%120%130%140%150%
Performance imprv EDP imprv
Original CATS CATA CATA+RSU
32-core system with 16 fast cores
25
Teaching opportunities (1)
Two courses in Computer Science BSc @ FIB (UPC)– PAR – Parallelism: 5th semester, mandatory course– PAP – Parallel Architectures and Programming: optional, computer
engineering specialization
OpenMP task-based programming @ node level
26
Teaching opportunities (2)Syllabus• Tasks, task graph, parallelism• Overheads and model: task creation, synchronization, data sharing• Task decomposition: linear, iterative, recursive• Task ordering and data sharing constraints• The magic of the architecture: cache coherency and synchronization
support• Data decomposition and influence of memory
OpenMP tasks at its heart, going towards task-only parallel programming • Embarrassingly parallel, recursive
task decomposition, task/data decomposition, and branch and bound
• Tareador tool for the exploration of task decomposition strategies
PAP
PAR
Innovation: (pseudo-) flipped-classroom model
27
Teaching opportunities (2)
PAP
PAR
Syllabus• Compiler and runtime support to OpenMP• Pthreads API and lib intrinsics• Designing an HPC cluster: components
and interconnect, roofline model for flops/bw• Message-passing interface MPI
OpenMP for node programming, with MPI towards cluster programming • OpenMP tasking review• Implementing a runtime to support OpenMP• Building a 4xOdroid cluster: hard and soft installation:
the effect of heterogeneity• Programming the 4xOdroid cluster
with MPI: Stencil with Jacobi
Innovation: puzzle work in groups when designing cluster
28
Thanks!
https://www.bsc.es